sign_extractor.py

#Extracting signature from input image

import cv2
import matplotlib.pyplot as plt
from skimage import measure, morphology
from skimage.measure import regionprops
import numpy as np


# the parameters are used to remove small size connected pixels outliar 
constant_parameter_1 = 84
constant_parameter_2 = 250
constant_parameter_3 = 100

# the parameter is used to remove big size connected pixels outliar
constant_parameter_4 = 18

def extract_sign(src_img):

    img = src_img
    img = cv2.threshold(img, 127, 255, cv2.THRESH_BINARY)[1]  # ensure binary

    # Connected component analysis which finds regions of connected pixels which have the same values

    blobs = img > img.mean()
    blobs_labels = measure.label(blobs, background=1)

    # to look into biggest component (having largest connceted pixesls with same values)
    the_biggest_component = 0
    total_area = 0
    counter = 0
    average = 0.0
    for region in regionprops(blobs_labels):
        if (region.area > 10):
            total_area = total_area + region.area
            counter = counter + 1
       
        if (region.area >= 250):
            if (region.area > the_biggest_component):
                the_biggest_component = region.area

    average = (total_area/counter)
    print("the_biggest_component: " + str(the_biggest_component))
    print("average: " + str(average))

    #threshold value to remove connected outliar connected pixels are smaller than a4_small_size_outliar_constant for A4 size scanned documents
    a4_small_size_outliar_constant = ((average/constant_parameter_1)*constant_parameter_2)+constant_parameter_3
    print("a4_small_size_outliar_constant: " + str(a4_small_size_outliar_constant))

    # remove the connected pixels are smaller than a4_small_size_outliar_constant
    pre_version = morphology.remove_small_objects(blobs_labels, a4_small_size_outliar_constant)

    #threshold value to remove outliar connected pixels are bigger than a4_big_size_outliar_constant for A4 size scanned documents
    a4_big_size_outliar_constant = a4_small_size_outliar_constant*constant_parameter_4
    print("a4_big_size_outliar_constant: " + str(a4_big_size_outliar_constant))


    # remove the connected pixels are bigger than threshold a4_big_size_outliar_constant 
    # to get rid of undesired connected pixels such as table headers and etc.
    component_sizes = np.bincount(pre_version.ravel()) #counting numbers of components whithin different size of them
    too_small = component_sizes > (a4_big_size_outliar_constant)
    too_small_mask = too_small[pre_version]
    pre_version[too_small_mask] = 0 # setting all smaller components as 0

    # save the the pre-version which is the image is labelled with colors as considering connected components
    plt.imsave('pre_version.png', pre_version)

    # read the pre-version
    img = cv2.imread('pre_version.png', 0)
    # ensure binary
    img = cv2.threshold(img, 0, 255,
                        cv2.THRESH_BINARY_INV | cv2.THRESH_OTSU)[1] #sign as Black colour and backgound as White 
    
    return img